摘要
针对皮肤病图像边界模糊且分布不规则、传统卷积分割方法无法满足对边缘细节提取的问题,提出了多级分裂卷积HSC-Net的皮肤病图像分割方法。网络编码端使用ImageNet上的VGG16-BN预训练模型,预训练参数会在训练过程中进行自动微调。将预训练模型中传统的最大池化层用软池化(Soft-pool)层进行替换,以减少传统池化的精度损失。解码端的HSC通过对特征图信息的分级提取,能高效利用特征信息。在解码端融入极化自注意力(Polarized Self-Attention, PSA)机制,使得空间和通道维度上获取更丰富的梯度信息。在ISIC2018数据集上的实验结果显示,精确度、Jaccard指数和Dice指数分别为96.21%、81.88%、81.65%,在准确性、轻量化和边界分割效果上优于现有的分割方法。
For the problem that the boundary of dermatological images is blurred and irregularly distributed and the traditional convolutional segmentation method cannot satisfy the extraction of edge details,a Hierarchical Split Convolution Network(HSC-Net)method for dermatological image segmentation is proposed.The network coding side uses the VGG16-BN pre-training model on ImageNet,and the pre-training parameters are automatically fine-tuned during the training process.Soft-pool layer replaces the traditional maximum pooling layer in the pre-training model to reduce the accuracy loss of traditional pooling.The decoding side uses Hierarchical Split Convolution(HSC)to efficiently utilize feature information.The decoder part uses the Polarized Self-Attention(PSA)mechanism,which enables the acquisition of richer gradient information in the spatial and channel dimensions.The experimental results on the ISIC2018 dataset show that accuracy,Jaccard index,and Dice index are 96.21%,81.88%,and 81.65%,respectively,which outperform existing segmentation methods in terms of accuracy,lightness and boundary segmentation effects.
作者
杨国亮
李林森
黄聪
黄经纬
YANG Guoliang;LI Linsen;HUANG Cong;HUANG Jingwei(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341411,China)
出处
《无线电工程》
北大核心
2023年第4期918-924,共7页
Radio Engineering
基金
国家自然科学基金(51365017)
江西省教育厅科技计划项目(GJJ190450,GJJ180484)。
关键词
图像处理
图像分割
多尺度特征融合
软池化
注意力机制
image processing
image segmentation
multiscale feature fusion
soft-pool
attention mechanism